Predicting Criminal Recidivism: A Comparison of Neural Network Models With Statistical Methods

• NCJ Number: 162895
• Journal: Journal of Criminal Justice Volume: 24 Issue: 3 Dated: (1996) Pages: 227-240
• Author(s): J Caulkins; J Cohen; W Gorr; J Wei
• Date Published: 1996
• Length: 14 pages
• Grant Number(s): 86-IJ-CX-0039
• Agencies: NIJ-Sponsored
• Publication Type: Report (Study/Research)

Annotation

This article applies neural network and conventional statistical models to predicting criminal recidivism.

Abstract

Neural networks can be viewed as nonlinear multiple regression models that use a new class of nonlinear forms. Parameter estimates for neural networks use the familiar method of minimizing the sum of squared errors, but invoke nonlinear search procedures. The primary innovation of neural networks is their family of nonlinear model forms, which provide two new pattern recognition capabilities. Neural network models provide substantial flexibility in which much of the model structure itself is estimated empirically from patterns recognized in the data. This capability may be useful in situations where there are inadequate theories for full model specification, rich collections on independent variables with complex interactions, subtle nonlinearities, or distinct submodels of unique behaviors. The data used for this research are from Gottfredson and Gottfredson (1979, 1980, 1985). These data include observations on recidivism during a 2-year period following release from Federal prisons for 3,508 offenders. The findings do not show any gains in accuracy by using neural networks to predict recidivism. Follow-up diagnostics reported in this study suggest that the variables used in available prediction measures do not have sufficient information to discriminate recidivists, although nonrecidivists can be better identified. Further improvements in predicting recidivism are likely to remain small without attention to theory building to better identify the behavioral mechanisms and contextual influences that affect recidivism and to develop improved measures of the resulting predictive factors. 8 tables, 1 figure, 52 references, and appended supplementary information on neural network models